Used RTX 3090 buying guide for local AI

Why people still buy it in 2026: 24 GB VRAM and 936 GB/s bandwidth (per llamabench.ai), the same VRAM as a 4090. New stock has mostly gone, so it's a used-market card.

What it actually does with 2026 models

Build	Model	Quant / setup	tok/s
4x RTX 3090 + 32-core Threadripper, 192 GB DDR4	Qwen3 235B-A22B	UD-Q3_K_XL	12-14
4x RTX 3090 + EPYC 7443, 256 GB DDR4 8-channel	Qwen3 235B-A22B	IQ4_XS	18
EPYC + single RTX 3090 (hybrid)	DeepSeek R1 671B	Q2_K	8.97
Single RTX 3090 24GB	gpt-oss 20B	MXFP4	161.77
RTX 3090: 2026 models (llamabench community)	Qwen3.6 35B-A3B	llama-bench tg128 (quant per result)	146-153
RTX 3090: 2026 models (llamabench community)	Qwen3.6 27B	llama-bench tg128 (quant per result)	40.2
RTX 3090: 2026 models (llamabench community)	Qwen3.8 27B	llama-bench tg128 (quant per result)	38.2-71.6
Single RTX 3090: Qwen3.6 35B-A3B (Giles Thomas)	Qwen3.6 35B-A3B	Q4 GGUF, all layers on GPU, ~90K ctx	139.6
4x RTX 3090 lab: Qwen3.6 on vLLM vs llama.cpp	Qwen3.6 35B-A3B	IQ4_XS (1 GPU)	115-133
4x RTX 3090 lab: Qwen3.6 on vLLM vs llama.cpp	Qwen3.6 35B-A3B	AWQ-INT4 (TP=2)	149
4x RTX 3090 lab: Qwen3.6 on vLLM vs llama.cpp	Qwen3.6 35B-A3B	AWQ-INT4 + MTP-3 (TP=2)	225-231
Single RTX 3090: Qwen3.8 27B at 78-140 tok/s with the OpenJet megakernel vs llama.cpp	Qwen3.8 27B	Q4_K_M (unsloth), greedy code prompt	139.7
Single RTX 3090: Qwen3.8 27B at 78-140 tok/s with the OpenJet megakernel vs llama.cpp	Qwen3.8 27B	Q4_K_M (unsloth), greedy code prompt	72.5
Single RTX 3090: Qwen3.8 27B at 78-140 tok/s with the OpenJet megakernel vs llama.cpp	Qwen3.8 27B	Q4_K_M (unsloth), reasoning mode	78.0
Single RTX 3090: Qwen3.8 27B at 78-140 tok/s with the OpenJet megakernel vs llama.cpp	Qwen3.8 27B	Q4_K_M (unsloth), reasoning mode	50.9
Single RTX 3090: Qwen3.8 27B at 78-140 tok/s with the OpenJet megakernel vs llama.cpp	Qwen3.8 27B	Q4_K_M (unsloth), no speculation	40.4
Single RTX 3090: Qwen3.8 27B at 78-140 tok/s with the OpenJet megakernel vs llama.cpp	Qwen3.8 27B	Q4_K_M (unsloth), no speculation	42.9
EPYC 7532 workstation (RTX Pro 6000 + RTX 3090 Ti, 300 W caps): Qwen3.8-Flash-Next, llama.cpp vs Strata engine	Qwen3.8 Flash-Next	IQ3_S, all experts on the RTX Pro 6000, no draft head	102.889
EPYC 7532 workstation (RTX Pro 6000 + RTX 3090 Ti, 300 W caps): Qwen3.8-Flash-Next, llama.cpp vs Strata engine	Qwen3.8 Flash-Next	IQ3_S + model draft head, all experts on the RTX Pro 6000	205.263
EPYC 7532 workstation (RTX Pro 6000 + RTX 3090 Ti, 300 W caps): Qwen3.8-Flash-Next, llama.cpp vs Strata engine	Qwen3.8 Flash-Next	IQ3_S, 32K prompt, all on the RTX Pro 6000	76.337
EPYC 7532 workstation (RTX Pro 6000 + RTX 3090 Ti, 300 W caps): Qwen3.8-Flash-Next, llama.cpp vs Strata engine	Qwen3.8 Flash-Next	IQ3_S, experts of 34/48 layers in system RAM, no draft head	22.657
EPYC 7532 workstation (RTX Pro 6000 + RTX 3090 Ti, 300 W caps): Qwen3.8-Flash-Next, llama.cpp vs Strata engine	Qwen3.8 Flash-Next	IQ3_S + draft head, 8,144 of 24,576 experts on card	98.291
RTX 3090 (llm-speed signed runs, Oct 2026): gpt-oss 20B 207 tok/s, Qwen3.6 27B 43 tok/s	gpt-oss 20B	reported as Q4_K_M	207.2
RTX 3090 (llm-speed signed runs, Oct 2026): gpt-oss 20B 207 tok/s, Qwen3.6 27B 43 tok/s	Qwen3.6 27B	Q4_K_M	42.7
RTX 3090 (llm-speed signed runs, Oct 2026): gpt-oss 20B 207 tok/s, Qwen3.6 27B 43 tok/s	GLM 4.7 Flash	Q4_K_M (Threadripper PRO host)	134.6
RTX 3090 (llm-speed signed run, Oct 2026): Flint 40B-A2.5B Q3_K_XL at 128.9 tok/s	Flint 40B-A2.5B	UD-Q3_K_XL	128.9
8x RTX 3090 serving DeepSeek V4 Flash (LimeChain vLLM kit)	DeepSeek V4 Flash	see source (FP8 KV cache)	95.52

Checklist before you pay

Ask for a screenshot of GPU-Z showing 24 GB GDDR6X and the serial matching the box/sticker.
Test in person if you can: run a few minutes of a GPU stress test or a llama.cpp benchmark and watch memory-junction temperatures (GDDR6X runs hot; check the card has working thermal pads).
Check fans, bent fins, missing screws and the 12-pin/8-pin connectors for discoloration.
Prefer buyer-protected platforms (eBay, Kleinanzeigen with protection) over cash transfers.
Plan for power: 3090 boards have high transient spikes; builders here often power-limit them (e.g. 270 W each).
Two 3090s (48 GB) give 70B-class and big MoE headroom, see the 4x3090 vLLM lab.

Search listings: eBay.de used RTX 3090. We don't quote a used price until we have a dated, sourced sample.

Not enough VRAM? Rent before you buy. Run a 3090 on a rented 24–96 GB cloud GPU by the hour and see the real speed before spending €1,000+. Try it on Runpod
Referral link: if you sign up with Google and add $10, Runpod gives you and RigBase bonus credit (terms). It never affects our data.

Compare: 3090 vs 4090 · best GPU 2026

New: Used RTX 3090 vs RTX 5090 · Best GPU under €1,000